基于说话人分类技术的分级说话人识别研究  

Research on Hierarchical Speaker Recognition Based on Speaker Clustering Technology

作　　者：刘文举[1] 孙兵[1] 钟秋海[2] 

机构地区：[1]中科院自动化所模式识别国家重点实验室,北京100080 [2]北京理工大学自动控制系,北京100081

出　　处：《电子学报》2005年第7期1230-1233,共4页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60172055;No.60121302);北京市自然科学基金(No.4042025).

摘　　要：识别正确率和抗噪性能固然是说话人识别的研究重点,但识别响应速度也是决定系统实用化的关键所在.本文成功地提出了基于说话人分类技术的分级说话人辨识方法,极大地提高了系统运行速度,随着注册说话人数的增多,较之传统的说话人辨识方法,其优势更加明显.同时在说话人确认中,该方法的使用,进一步提高了确认的正确率,有效地降低了错误接受和错误拒绝率.本文提出的可信度打分方法,也一定程度上改进了系统的性能.实验表明:基于说话人分类技术的说话人辨识方法使系统的运行速度平均提高了3.5倍,对说话人确认等误识率和最小误识率平均下降了53.75%.Recognition correct rate and noise robust property are indeed important for speaker recognition research,but the respense rate of recognition is also a key factor for a speaker recognition system when applied in the real world. Owing to this, we propose a novel speaker identification approach based on speaker clustering, namely Hierarchical Speaker Identification （HSI）. It can increase the nmning speed greatly for speaker identification systems,and the more the number of registered speakers is,the faster the HSI system runs than the Conventional Speaker Identification （CSI） system. Simultaneously, its counterpart for speaker verification based on speaker clustering, can reduce the rates of false rejection and false acceptance efficiently to improve the capability of verification. A new method is also presented here called reliability scoring. The experiments show that speaker clustering based algorithms can run faster 3.5 times than original approach for the speaker identification and is 53.75% deduction of equal or minimal error rates for the speaker verification on average.

关 键 词：说话人辨识 说话人确认 说话人分类 Cohort集 可信度打分 

分 类 号：TN912.34[电子电信—通信与信息系统]